Leaders Insights
Leaders Insights

Rester au meilleur niveau, un peu chaque jour.

DomainesMarketingDataFinanceIA
RessourcesApprendreTestOutilsBlogGlossaire
© 2026 Leaders Insights — Tous droits réservés.
Formations/AI in asset management/Use cases, ROI and evaluation/Mapping AI across the asset management value chain
1/5+150 XP

Use cases, ROI and evaluation

5Mapping AI across the asset management value chain+1506Separating genuine AI use cases from vendor theater+1507Evaluating and running proof-of-concepts with AI vendors+1508Building a defensible ROI model for AI initiatives+1509Setting realistic expectations and adoption roadmaps+150

Mapping AI across the asset management value chain

# Mapping AI across the asset management value chain

A large European asset manager recently disclosed that its analysts were spending roughly a third of their week reading filings, broker notes, and earnings transcripts before writing a single line of investment thinking. That reading pile is where AI first earned its keep in this industry: not by picking stocks, but by compressing the hours between raw information and a usable insight.

This lesson walks the full pipelinepipelineAll active sales opportunities across the stages of the sales process, together with their combined potential value and probability of closing.Voir la définition complète →, from research through back-office operations, and marks each stage as real edge, useful but modest, or mostly hype for 2026.

The value chain, stage by stage

Think of an asset management firm (a company that invests money on behalf of clients: pension funds, insurers, individuals) as a sequence of steps. Money and information flow through each one. AI attaches differently at each stage.

1. Research and idea generation

Verdict: real edge.

This is where large language models (LLMs, AI systems trained to process and generate text) deliver the clearest wins today. The value is not exotic. It is speed and coverage.

Concrete uses:

  • Summarizing 10-KKThe average number of new users each existing user generates through referrals. Above 1.0, growth compounds on itself and becomes exponential.Voir la définition complète → and 10-Q filings (annual and quarterly reports US companies file with the Securities and Exchange Commission, the SEC) and flagging changes in risk language year over year.
  • Transcribing and querying earnings calls: "Did management soften guidance on margins?"
  • Scanning thousands of documents an analyst would never have time to read.

The honest limit: LLMs hallucinate (generate confident but false statements). No serious desk lets a model's summary reachreachThe number of unique people exposed to your message in a given period. Unlike impressions, reach counts each person once, no matter how often they see it.Voir la définition complète → a portfolio manager without a source link back to the original sentence. The workflow is "AI drafts, human verifies," not "AI decides."

2. Quantitative signals and alpha models

Verdict: real edge, but not new.

Quant funds (firms using statistical models to trade) have used machine learning for over a decade. Two Sigma, Renaissance, and AQR were doing this long before ChatGPT. What is new in 2026 is alternative dataalternative dataDonnées non-traditionnelles utilisées pour l'analyse d'investissement ou le renseignement concurrentiel : images satellites, transactions bancaires, géolocalisation, scraping web, mentions sociales. processed by AI: satellite imagery of parking lots, credit card aggregates, shipping data, and sentiment extracted from text.

Be skeptical of overclaims. Markets are adaptive: a signal that works stops working once enough people trade it. AI does not repeal this. It lowers the cost of finding signals, which means signals decay faster too.

3. Manager selection and due diligence

Verdict: useful but modest.

Firms that allocate to external managers (funds of funds, wealth platforms) use AI to parse due diligence questionnaires, cross-check disclosures, and spot inconsistencies across documents. Useful for catching red flags at scale. It does not replace the qualitative judgment of whether a manager's edge is real or luck.

4. Portfolio construction and risk

Verdict: useful but modest.

Optimization math here is mature and does not need generative AI. Where AI helps: scenario generation ("show me portfolios stressed against a 2022-style rate shock") and natural-language risk queries so a non-quant PM can ask questions without an analyst intermediary. Genuine convenience. Not a source of alpha by itself.

5. Trade execution

Verdict: real edge, quietly.

Execution algorithms that decide how to slice a large order to minimize market impact (the price move your own trading causes) have used reinforcement learning for years. This is unglamorous and genuinely effective. Better execution saves basis points on every trade, and those compound.

6. Client reporting and servicing

Verdict: real edge on cost, watch compliance.

Generating commentary for hundreds of client portfolios used to be manual. LLMs now draft first-pass quarterly letters and answer routine client questions. The catch: anything client-facing is regulated communication. In the US the SEC's marketing rule governs it; in the EU, MiFID II (Markets in Financial Instruments Directive, the rulebook for investment services) does. A hallucinated performance figure in a client letter is a compliance incident, not a typo. Human sign-off is mandatory in practice.

7. Back-office operations

Verdict: real edge, underrated.

The least glamorous stage may offer the best ROIROIReturn on Investment: the ratio of net profit to the cost of an investment. A 300% ROI means each dollar invested returns $3.Voir la définition complète →. Reconciliation, trade matching, corporate actions processing, and Know Your Customer (KYC) checks are high-volume, rules-heavy, and error-prone. AI plus document extraction cuts manual handling. Because these costs are large and boring, automating them rarely gets headlines but frequently pays back fastest.

A simple ROIROIReturn on Investment: the ratio of net profit to the cost of an investment. A 300% ROI means each dollar invested returns $3.Voir la définition complète → lens

Do not evaluate AI on "is it impressive." Evaluate on time saved times cost of that time, minus the cost to run and govern the tool.

Worked example (illustrative figures, not a benchmark):

Research summarization tool
- Analysts affected:        20
- Hours saved per week:     5 each
- Blended cost per hour:    $150 (illustrative)
- Weekly saving:            20 x 5 x $150 = $15,000
- Annual saving (46 wks):   $690,000

Annual cost to run:
- Software + LLM API:       $180,000 (illustrative)
- Oversight / review time:  $120,000 (illustrative)
- Total cost:               $300,000

Net annual benefit:         $690,000 - $300,000 = $390,000

The point is the discipline, not the number. Notice the oversight cost line. Teams that forget it overstate ROIROIReturn on Investment: the ratio of net profit to the cost of an investment. A 300% ROI means each dollar invested returns $3.Voir la définition complète →. The verification burden is a permanent operating cost, not a one-off.

For a grounded view of where these tools actually help versus where risk sits, the Bank of England and FCA machine learning surveys are a useful free reference on adoption patterns in financial services.

🎬 [VIDEO: "How AI Is Actually Used in Asset Management" - https://www.youtube.com/results?search_query=artificial+intelligence+asset+management - practitioner overview of AI applications across investment workflows]

Where the hype concentrates

Three claims to challenge whenever you hear them:

1. "Our AI predicts the market." No public evidence supports durable, generalizable market prediction from generative AI. Treat this as a marketing red flag.

2. "Fully autonomous portfolio management." Regulators require accountable humans. Under MiFID II and SEC rules, someone must be responsible for decisions. Full autonomy is a governance non-starter in 2026, regardless of technical feasibility.

3. "It replaces analysts." It changes what analysts do (less reading, more judgment), and augments rather than replaces. Firms that framed it as headcount replacement have generally walked that back.

Vérification des acquis

1. According to the lesson, why does research and idea generation earn a 'real edge' verdict for AI in asset management?

2. What does the 'AI drafts, human verifies' workflow primarily guard against?

3. Why does the lesson label quantitative signals and alpha models as 'real edge, but not new'?

CHOIX MULTIPLES

4. Select ALL correct answers about legitimate uses of LLMs in the research stage as described in the lesson.

Sélectionnez toutes les réponses correctes.

CHOIX MULTIPLES

5. Select ALL correct answers describing the framework this lesson uses to evaluate AI across the value chain.

Sélectionnez toutes les réponses correctes.

How to evaluate an AI solution in this sector

When a vendor pitches you, run five checks:

Traceability. Can every AI output link back to a verifiable source? If a summary cannot point to the filing sentence it came from, it is not usable in a regulated shop.

Data boundary. Where does your data go? A manager feeding non-public research into a third-party model must know whether that data trains the vendor's model. Contractually, most institutional deals now forbid this. Confirm it.

Model risk governance. In the US, model risk guidance (originating from the Federal Reserve and OCC, framework SR 11-7) increasingly gets applied to AI tools. Ask how the vendor supports validation, monitoring, and documentation.

Failure behavior. What happens when the model is uncertain? A good tool says "I do not know" or flags low confidence. A bad one always answers confidently. The second is dangerous in finance.

Measurable outcome. Insist on a metric before purchase: hours saved, error rate reduced, coverage expanded. "It feels faster" is not an evaluation.

Adoption realities

The technology is rarely the bottleneck. The bottlenecks are:

  • Data plumbing. Firms with messy, siloed data cannot deploy AI well. Cleanup is 60 to 80 percent of many projects (widely cited industry estimate, varies by firm).
  • Compliance sign-off. Legal and compliance review adds months. Build it into timelines.
  • Change management. PMs and analysts must trust the tool. Trust comes from transparency (traceability again), not from accuracy claims.

Start where the risk is low and the volume is high: back-office reconciliation, research summarization, internal knowledge search. Prove ROIROIReturn on Investment: the ratio of net profit to the cost of an investment. A 300% ROI means each dollar invested returns $3.Voir la définition complète → there before touching anything client-facing or investment-decision-facing.

Key takeaways

  • AI edge is uneven across the chain. Strongest in research summarization, execution algorithms, and back-office automation. Weakest (and most hyped) in "market prediction" and autonomous decision-making.
  • Verification is a permanent cost. LLMs hallucinate, so every ROIROIReturn on Investment: the ratio of net profit to the cost of an investment. A 300% ROI means each dollar invested returns $3.Voir la définition complète → model must budget for human oversight. Tools that cannot trace outputs to sources are not deployable in regulated workflows.
  • Regulation sets hard limits. MiFID II, SEC marketing rules, and model risk frameworks (SR 11-7) mean an accountable human stays in the loop. Full autonomy is off the table for 2026.
  • Evaluate on measurable outcomes, not impressiveness. Time saved, error rates cut, coverage expanded. Demand a metric before you buy.
  • Sequence adoption by risk. Win in high-volume, low-risk back-office and research tasks first. Earn the right to move toward client and investment decisions later.

Suivant

Separating genuine AI use cases from vendor theater